Buckets:
| # cwm3/eval/run_gaze_evals.sh — evaluation of the gaze-chain U0 runs (gaze/robustness agent, 2026-09-12; results/tables/gaze_chain.md), CPU. | |
| # bash eval/run_gaze_evals.sh "<seeds>" [render=1] | |
| # Per seed k and regime (all tags idempotent on results/paired_compare.csv; driver populations from the v2 causal kin targets): | |
| # gaze_s<k>_gz_vs_clip8ln / gaze_s<k>_gzfc_vs_clip8ln A = U0_kin_can_noK_s<k>_kcv2_clip8_ln_{gz,gzfc} B = …_kcv2_clip8_ln (survivor) — K4-style driver cells, identical rows | |
| # gaze_s<k>_gzfc_vs_gz A = _gzfc B = _gz — the face-input effect | |
| # gaze_s<k>_{gz,gzfc}_vs_k0 B = k0 heuristic (+ persistence base-rate rows) — levels vs the label-free references | |
| # gazeon_s<k>_{gz,gzfc}_vs_rule eval/gaze_chain_eval.py: gaze_off_on / off @3s PR-AUC (+ AUROC) vs the current-state rule |lm_yaw_dev|(t) | |
| # gazeon_s<k>_gzfc_vs_gz, gazeon_s<k>_gz_persist the face effect on the gaze chain; persistence base rate | |
| # gazegold_s<k>_{gz,gzfc} --dms-gold: gold clean-test rows, our gaze_off_on label, model vs DMS att@3s / rule / persistence -> results/gaze_chain/gaze_gold.csv | |
| set -u | |
| cd "$(dirname "$0")/.." | |
| PY=python | |
| export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4 | |
| have_tag() { grep -q ",$1,$2,test," results/paired_compare.csv 2>/dev/null; } | |
| have_gold() { grep -q ",$1,$2," results/gaze_chain/gaze_gold.csv 2>/dev/null; } | |
| done_run() { [[ -f results/runs/$1/PRED_DONE && -f results/runs/$1/TRAIN_DONE ]]; } | |
| pc() { # tag a b [extra] | |
| local tag="$1" a="$2" b="$3"; shift 3 | |
| done_run "$a" || { echo "skip $tag ($a not finished)"; return; } | |
| [[ "$b" == none ]] || done_run "$b" || { echo "skip $tag ($b not finished)"; return; } | |
| for reg in manual engaged; do | |
| have_tag "$tag" "$reg" && continue | |
| echo "== $tag $reg" | |
| $PY eval/paired_compare_joint.py --regime "$reg" --a "results/runs/$a" --b "results/runs/$b" --tag "$tag" --kin-targets causal_routes --kin-version v2 "$@" 2>&1 | grep -E "appended|Error|Traceback|File \"|LAYOUT GUARD" | |
| done | |
| } | |
| ge() { # tag a [extra] (gaze_chain_eval population mode) | |
| local tag="$1" a="$2"; shift 2 | |
| done_run "$a" || { echo "skip $tag ($a not finished)"; return; } | |
| for reg in manual engaged; do | |
| have_tag "$tag" "$reg" && continue | |
| echo "== $tag $reg" | |
| $PY eval/gaze_chain_eval.py --regime "$reg" --a "results/runs/$a" --tag "$tag" "$@" 2>&1 | grep -E "appended|Error|Traceback|File \"|LAYOUT GUARD|gaze_onset" | |
| done | |
| } | |
| for s in ${1:-0 1 2}; do | |
| B=U0_kin_can_noK_s${s}_kcv2_clip8_ln; GZ=${B}_gz; FC=${B}_gzfc | |
| pc "gaze_s${s}_gz_vs_clip8ln" $GZ $B | |
| pc "gaze_s${s}_gzfc_vs_clip8ln" $FC $B | |
| pc "gaze_s${s}_gzfc_vs_gz" $FC $GZ | |
| pc "gaze_s${s}_gz_vs_k0" $GZ none --b-k0 | |
| pc "gaze_s${s}_gzfc_vs_k0" $FC none --b-k0 | |
| ge "gazeon_s${s}_gz_vs_rule" $GZ --b-rule --auroc | |
| ge "gazeon_s${s}_gzfc_vs_rule" $FC --b-rule --auroc | |
| ge "gazeon_s${s}_gzfc_vs_gz" $FC --b "results/runs/$GZ" --auroc | |
| ge "gazeon_s${s}_gz_persist" $GZ --b-persistence | |
| for v in gz gzfc; do | |
| d=${B}_${v}; tag=gazegold_s${s}_${v} | |
| done_run $d || { echo "skip $tag"; continue; } | |
| have_gold "$tag" "$s" && continue | |
| echo "== $tag" | |
| $PY eval/gaze_chain_eval.py --dms-gold --a "results/runs/$d" --seed $s --tag "$tag" --n-boot 2000 2>&1 | grep -E "appended|Error|Traceback|File \"|rows in the|gold clean_test" | |
| done | |
| done | |
| if [[ "${2:-1}" == "1" ]]; then $PY eval/report_gaze_chain.py --seeds ${1:-0 1 2} --out results/tables/gaze_chain.md 2>&1 | tail -3; fi | |
| echo GAZE_EVALS_DONE | |
Xet Storage Details
- Size:
- 3.68 kB
- Xet hash:
- de4e27f85a579c4cfec6063e968588435e30c0235272d592ad9f8d09a92f64bf
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.